Towards Efficient Fault Detection of Ultra-High Voltage Direct Current Circuit Breakers

Fault detection on Ultra-High Voltage (UHV) Direct Current (DC) circuit breakers is crucial for the safety and reliability of electrical systems. Existing fault detection algorithms leverage deep neural networks (DNNs) to achieve high detection accuracy under various conditions. However, these methods often introduce considerable delays in fault detection, failing to meet the stringent time requirements for fault detection of UHV DC breakers. To resolve this issue, we propose a training and pruning framework designed to accelerate DNN-based detection models without compromising accuracy. Our framework treats the training of the detection model as a multi-objective optimization problem and utilizes the alternating direction method of multipliers (ADMM) to simultaneously train and prune the detection model. Moreover, since UHV DC breakers are high-value devices, it is infeasible to gather large amounts of fault case data. Therefore, we propose a self-supervised learning module for the proposed framework to pretrain the detection model using normal case data and finetune it using a small amount of fault case data. Experimental results demonstrate that the detection model trained with our framework surpasses baseline models in both mean average precision (98.7%) and inference latency (4ms), providing a more efficient and accurate solution for UHV DC circuit breaker fault detection.